J. R. Wallis’s research while affiliated with University of Washington and other places

Spatial patterns in trends of four monthly variables: average temperature, precipitation, streamflow, and average of the daily temperature range were examined for the continental United States for the period 1948-88. The data used are a subset of the Historical Climatology Network (1036 stations) and a stream gage network of 1009 stations. Trend significance was determined using the nonparametric seasonal Kendall's test on a monthly and annual basis, and a robust slope estimator was used for determination of trend magnitudes. A bivariate test was used for evaluation of relative changes in the variables, specifically, streamflow relative to precipitation, streamflow relative to temperature, and precipitation relative to temperature.Strong trends were found in all of the variables at many more stations than would be expected due to chance. There is a strong spatial and seasonal structure in the trend results. For instance, although annual temperature increases were found at many stations, mostly in the North and West, there were almost as many downtrends, especially in the South and East. Among the most important trend patterns are (a) increases in March temperature at almost half of the stations; (b) increases in precipitation from September through December at as many as 25 percent of the stations, mostly in the central part of the country; (c) strong increases in streamflow in the period November-April at a maximum of almost half of the stations, with the largest trend magnitudes in the north-central states; (d) changes in the temperature range (mostly downward) at a large number of stations beginning in late spring and continuing through winter, affecting as many as over half of the stations. The observed trends in streamflow are not entirely consistent with the changes in the climatic variables and may be due to a combination of climatic and water management effects.

Spatial patterns in trends of four monthly variables: average temperature, precipitation, streamflow, and average of the daily temperature range were examined for the continental United States for the period 1948-88. The data used are a subset of the Historical Climatology Network (1036 stations) and a stream gage network of 1009 stations. Trend significance was determined using the nonparametric seasonal Kendall's test on a monthly and annual basis, and a robust slope estimator was used for determination of trend magnitudes. A bivariate test was used for evaluation of relative changes in the variables, specifically, streamflow relative to precipitation, streamflow relative to temperature, and precipitation relative to temperature. Strong trends were found in all of the variables at many more stations than would be expected due to chance. There is a strong spatial and seasonal structure in the trend results. For instance, although annual temperature increases were found at many stations, mostly in the North and West, there were almost as many downtrends, especially in the South and East. Among the most important trend patterns are (a) increases in March temperature at almost half of the stations; (b) increases in precipitation from September through December at as many as 25 percent of the stations; (c) strong increase in streamflow in the period November-April at a maximum of almost half of the stations, with the largest trend magnitudes in the north-central states; (d) changes in the temperature range (mostly downward) at a large number of stations beginning in late spring and continuing through winter, affecting as many as over half of the stations. The observed trends in streamflow are not entirely consistent with the changes in the climatic variables and may be due to combination of climatic and water management effects. 34 refs., 16 figs, 7 tabs.

Previous attempts to validate general circulation model simulations of land surface hydrology have often been limited by the absence of systematic historical data, especially for runoff, precipitation, and temperature. Because hydrological response times for unregulated watersheds in the United States vary from a few hours to a few days at most, climatological studies dealing with land surface hydrology require data at relatively short time intervals. The authors describe a set of 1,009 US Geological Survey streamflow stations, and 1,036 National Oceanic and Atmospheric Administration climatological stations, for which long-term (1948-1988) observations have been assembled into a consistent daily data base with missing observations estimated using a simple closest-station prorating rule. Care was taken in selection of the streamflow stations to assure that the records were essentially free from regulation. The climatological stations are a subset of the historical climatology network for which development of alternative data retrieval algorithms. estimated values for missing data, as well as suspicious observations, are flagged. The data are retrievable by station list, state, latitude-longitude range, and hydrologic unit code from compact digital read-only memory (CD ROM). CD-ROM copies are available from D.P. Lettenmaier.

Regional flood frequency analysis usually assumes that flood records from different sites are statistically independent. This assumption is unlikely to be valid in practice, so it is important to know how intersite dependence affects flood quantile estimates obtained by regional analysis. We use Monte Carlo simulation to assess the effect of realistically specified intersite dependence on the regional probability weighted moment algorithm, a robust and efficient procedure for regional flood frequency analysis. Our principal conclusions are as follows. Any bias in flood quantile estimates is unchanged by the presence of intersite dependence. The accuracy of flood quantile estimates decreases when intersite dependence is present, but this effect is less important for practical applications than the bias in flood quantile estimates due to heterogeneity (inequality of the flood frequency distributions in the region). Even when both heterogeneity and intersite dependence are present and the form of the flood frequency distribution is misspecified, regional flood frequency analysis is more accurate than at-site analysis.

The generalized Pareto distribution is a two-parameter distribution that contains uniform, exponential, and Pareto distributions as special cases. It has applications in a number of fields, including reliability studies and the analysis of environmental extreme events. Maximum likelihood estimation of the generalized Pareto distribution has previously been considered in the literature, but we show, using computer simulation, that, unless the sample size is 500 or more, estimators derived by the method of moments or the method of probability-weighted moments are more reliable. We also use computer simulation to assess the accuracy of confidence intervals for the parameters and quantiles of the generalized Pareto distribution.

Recent work on regional flood frequency estimation has shown that accurate flood quantile estimates are possible when the underlying flood frequency distributions are identical at all sites in the region except for a scaling factor, particularly when the underlying distribution has a two-parameter form. The class of regional probability-weighted moment (PWM) estimators is investigated for robustness to mis-specification of the assumed distributional form and to regional heterogeneity in moments of order higher than one. Whereas two-parameter distributions belonging to the extreme value family perform quite well when the form of the underlying distribution is close to that of the fitted distribution, large biases can result when the distribution is misspecified. The three-parameter generalized extreme value distribution (GEV), when fitted using the regional PWM method, has been shown to be relatively insensitive to violations of the distributional assumption, and to have low variability and bias. In this paper it is shown that regional estimation methods using the three-parameter GEV distribution are relatively insensitive to modest regional heterogeneity in the coefficient of variation and quite insensitive to regional variation in the skew coefficient. The key determinant of the performance of the regional estimators is shown to be the regional mean coefficient of variation. For high values of the mean coefficient of variation, such as might be encountered in arid regions, an alternate PWM estimation method based on the GEV distribution that accommodates the regional heterogeneity in the higher order moments is preferred. The trade-off between this alternate method and the approach that assumes regional homogeneity in moments higher than order one is sensitive to the record lengths.

A common problem in flood frequency analysis is estimation, from a short gauged record, of the flood corresponding to a return period of 50-1000 years. One would expect that an estimate of the historical maximum event; i.e., the maximum flood in a period of perhaps 100-200 years, would, if properly included in the estimation procedure, improve the resulting estimate of the flood magnitude at the given return period. We attempt to assess the value of historical information using computer simulation. An annual flood sequence and a historical maximum event are generated from ah extreme-value distribution (of type I or II) and quantitiles of the distribution are estimated by maximum likelihood, both including the excluding the historical event. After many such simulations have been performed the accuracy of the quantile estimates is assessed. We also consider the effect of errors in the magnitude of historical discharge estimates and the use of historical data in regional flood frequency analysis.

Techniques recently developed in paleohydrology enable the estimation of discharges and dates of occurrence of flood events which occurred thousands of years ago. We use computer simulation to assess whether a single paleoflood estimate, when included in a single-site or regional flood frequency analysis procedure, gives a worthwhile increase in the accuracy of estimates of extreme floods. We find that the main factors affecting the utility of this kind of paleological information are the specification of the fitted flood frequency distribution (whether it has two or three unknown parameters) and the size of the measurement error of paleodischarge estimates. Errors in estimating the date of the paleoflood are of minor importance. We conclude that paleological information is most useful when estimating a three-parameter flood frequency distribution for a single site possessing only a short gauged record. When several independent and homogeneous gauged records from different sites are used in a regional flood frequency analysis, the value of paleological information is, within the limits of this study, negligible.